Mark my words!

Citation

Cristian Danescu-Niculescu-Mizil, Michael Gamon, Susan T. Dumais: Mark my words!: linguistic style accommodation in social media. WWW 2011: 745-754

Online version

http://research.microsoft.com/en-us/um/people/sdumais/fr862-danescu-niculescu-mizil.pdf

Summary

Physiological studies have suggested that participants in conversation accommodate in dimensions such as speaking style, utterance length, gesture, speaking rate etc. In this paper authors proves the hypothesis that linguistic accommodation could be seen in social media such as Twitter. They investigate accommodation in LIWC dimensions[1]. Some examples of these dimensions include use of article,negation words(not/no),preposition,quantifier,1st person singular pronoun,1st person plural pronoun,2nd person pronoun in conversation. Authors propose a novel probabilistic framework to prove their hypothesis.

Framework

When individual talk about some topic, they would have to use similar words to describe topics hence it is important to remove topic accommodation from overall accommodation measure. Since they use LIWC dimensions, it is automatically removed.

Their framework is based on mainly two components, stylistic cohesion and stylistic accommodation.

Stylstic Cohesion: It is used to find if tweets belonging to same conversation exhibit a certain LIWC style more than tweets which are unrelated. If the former is more then we can say that tweets which are part of same conversation agree more on a particular style. Formally, for a style $C$ it is defined as:

$Coh(C)=P(T^{C}\wedge R^{C}|T\leftrightarrow R)-P(T^{C}\wedge R^{C})$

where $T\leftrightarrow R$ is condition which represent that tweets are from same conversation. $P(T^{C}\wedge R^{C}|T\leftrightarrow R)$ is the probability of tweets which are part of same conversation and exhibit style C. Whereas, $P(T^{C}\wedge R^{C})$ is probability of observing style C in any randomly picked two tweets.

Stylistic accommodation:

While measuring stylistic accommodation it is assumed that a twitter can accommodate in a style with his partner only if his partner exhibited style C in same conversation earlier. The formal definition of stylistic definition is as follows:

$Acc(a;b)^{C}=P(T_{b}^{C}|T_{a}^{C},T_{b}\rightarrow T_{a})-P(T_{b}^{C}|T_{b}\rightarrow T_{a})$

Here, $P(T_{b}^{C}|T_{a}^{C},T_{b}\rightarrow T_{a})$ represents the probability that style $C$ was exhibited in tweets of user b after observing the same style in user $a$. Whereas, $P(T_{b}^{C}|T_{b}\rightarrow T_{a})$ represent that style C was observed in user b irrespective of whether user a used the style C or not. Note the fact that $Acc(a;b)$ is directional accommodation from a to b. They also defines accommodation from b to a. They use these two accommodation scores $Acc(a;b)$ and $Acc(b;a)$ to find if accommodation is symmetric or not.

Results:

Authors observe that $P(T^{C}\wedge R^{C}|T\leftrightarrow R)$ is more than $P(T^{C}\wedge R^{C})$ in considered LIWC styles. This confirms the fact that Stylistic Cohesion is present in Twitter. They also observe that $P(T_{b}^{C}|T_{a}^{C},T_{b}\rightarrow T_{a})$ is more than $P(T_{b}^{C}|T_{b}\rightarrow T_{a})$. This confirms that linguistic accommodation in LIWC styles is present in twitter.

Related Paper:

• Rivka Levitan, Agustín Gravano, Julia Hirschberg: Entrainment in Speech Preceding Backchannels. ACL (Short Papers) 2011: 113-117
• Rivka Levitan, Julia Hirschberg: Measuring Acoustic-Prosodic Entrainment with Respect to Multiple Levels and Dimensions. INTERSPEECH 2011: 3081-3084